Glossary

Key terms and concepts used throughout this field guide, defined in the context of the biomarker discovery project.

A

Batch Effects
Non-biological variation in gene expression data arising from technical differences between experimental batches (e.g., different sequencing runs, library preparation dates, labs). In this project, study-specific effects consistently dominated trait effects, limiting the effectiveness of integrated analysis approaches.

Biomarker
A measurable molecular indicator (in this case, gene expression level) that distinguishes between biological states (tolerant vs. sensitive phenotypes). The project identified a 6-gene biomarker panel.

C

Classifier
A machine learning model that predicts categorical outcomes (phenotypes) based on input features (gene expression). This project used logistic regression with LASSO regularization to build a minimal 6-gene classifier.

COMBAT
A batch correction method that adjusts for systematic technical variation using empirical Bayes. Attempted in Issue #18 with limited success.

Cross-Validation
A technique for assessing how well a model generalizes to independent data by systematically holding out portions of data for testing. See: LOSO.

D

DEG (Differentially Expressed Gene)
A gene whose expression level differs significantly between two or more conditions (e.g., resistant vs. sensitive, control vs. treated).

DESeq2
A widely-used R package for differential gene expression analysis from RNA-seq count data. Used throughout this project for identifying DEGs.

Differential Abundance
Analysis method to identify features (genes, transcripts) that differ in quantity/expression between experimental groups. This project primarily used the nf-core/differentialabundance pipeline.

Directionality Consistency
In the two-step classifier approach, the requirement that a gene’s fold change direction (up or down) is the same across multiple datasets. Ensures reproducibility of effect direction.

F

FastP
A tool for quality control and preprocessing of FASTQ files (raw sequencing reads). Parameter selection is critical for different library types (e.g., TAG-seq vs. standard RNA-seq) - see Issue #26, #28.

Fold Change
The ratio of gene expression levels between two conditions, often expressed as log2 fold change. Positive values indicate upregulation, negative values indicate downregulation.

G

GMT File
Gene Matrix Transposed file format used for gene set definitions in GSEA. Contains pathway/gene set names with associated genes. Created in this project for GSEA integration.

GSEA (Gene Set Enrichment Analysis)
A computational method that determines whether a defined set of genes (e.g., a pathway) shows statistically significant differences between two biological states. Attempted in Issue #45.

H

Heterogeneity
In the context of gene scoring, the variance in gene expression within phenotype groups. Low heterogeneity (high within-group consistency) is desirable for biomarkers.

I

Innate Biomarker
A gene that is constitutively different in expression between resistant and sensitive individuals, even in the absence of stress. Contrasts with reactive biomarkers. Key insight from Issue #53.

Integrated Data Analysis
Approach that pools multiple datasets together before analysis, treating all samples as if from a single study (with batch correction). In this project, integrated analysis failed due to strong study-specific effects (Big Lesson #1).

L

LASSO (Least Absolute Shrinkage and Selection Operator)
A regularization method for regression that can shrink coefficients to exactly zero, performing automatic feature selection. Used in the two-step classifier to identify the minimal gene set.

Leakage (Data Leakage)
When information from the test set inappropriately influences the training process, leading to overly optimistic performance estimates. See Validation & Pitfalls.

LOSO (Leave-One-Study-Out)
A cross-validation strategy where each study is held out once as a test set, while all other studies form the training set. Critical for assessing cross-study generalization (Big Lesson #4).

M

Meta-Analysis
Statistical approach that combines results from multiple independent studies to identify consistent patterns. This project used a post-data integration meta-analysis approach.

Mutual Information
A measure of statistical dependence between variables. Used early in the project to assess gene-phenotype associations, but trait separation was weak.

N

nf-core
A community effort to collect curated bioinformatics pipelines built using Nextflow. This project used nf-core/rnaseq and nf-core/differentialabundance.

Normalization
Process of adjusting gene expression data to account for technical variation (library size, sequencing depth, GC content) while preserving biological variation. Critical for cross-sample comparison.

P

PCA (Principal Component Analysis)
A dimensionality reduction technique that identifies major sources of variation in high-dimensional data (like gene expression). Used throughout the project to visualize sample clustering and assess batch effects.

Perkinsus marinus
Protozoan parasite that causes Dermo disease in oysters, the primary stressor in this project’s datasets.

Phenotype
Observable trait or characteristic. In this project: tolerant/resistant vs. sensitive/susceptible oyster phenotypes in response to P. marinus infection.

Post-Data Integration
Approach that analyzes each dataset independently, then compares results across datasets to identify reproducible findings. This project pivoted to this approach after integrated analysis failed.

R

Reactive Biomarker
A gene whose differential expression between resistant and sensitive individuals emerges only after stress exposure (i.e., significant in treated samples but not controls). Contrasts with innate biomarkers.

RemoveBatchEffect
A function from the limma R package for adjusting gene expression data to remove batch effects. Attempted in Issue #18 with limited success.

Reproducibility
In the context of the two-step classifier, the degree to which a gene is identified as differentially expressed across multiple independent datasets. High reproducibility indicates robust, generalizable biomarkers.

RNA-seq
RNA sequencing - a high-throughput method for quantifying gene expression by sequencing cellular RNA. The primary data type in this project.

S

Stepwise Differential Abundance
A two-step filtering approach: (1) identify stress-responsive genes (control vs. treated), then (2) identify resistance-associated genes from Step 1 genes. Has limitations (removes innate biomarkers, breaks down with small gene sets) - see Stepwise Pipeline.

Study-Specific Effects
Systematic differences between datasets due to experimental design, protocols, or biological differences between populations. In this project, these effects were stronger than trait effects, driving the pivot to post-data integration.

T

TAG-seq (3’ Tag RNA-Sequencing)
A cost-effective alternative to standard RNA-seq that sequences only the 3’ end of transcripts. Requires different data processing parameters than full-length RNA-seq (see TAG-seq issues).

Tolerant/Resistant
Phenotype category for oysters that survived P. marinus infection or showed low infection intensity and minimal pathology.

Trait Effects
Gene expression differences attributable to the biological phenotype of interest (resistant vs. sensitive). In this project, trait effects were initially weaker than study-specific effects.

V

Validation
Process of confirming that a biomarker panel or model performs well on independent data not used during development. See Validation & Pitfalls.

VST (Variance-Stabilizing Transformation)
A DESeq2 normalization method that transforms count data to a scale where variance is roughly constant across the range of expression values. Enables visualization and clustering. Requires sufficient genes (>1000) for robust estimation.

Abbreviations

  • DE: Differential Expression
  • DEG: Differentially Expressed Gene
  • FDR: False Discovery Rate (adjusted p-value threshold)
  • GC Bias: Systematic bias related to GC content in sequencing
  • GTF: Gene Transfer Format (gene annotation file)
  • HPC: High-Performance Computing
  • LOSO: Leave-One-Study-Out
  • PCA: Principal Component Analysis
  • QC: Quality Control
  • WGBS: Whole-Genome Bisulfite Sequencing (for methylation)

Species

  • Crassostrea virginica: Eastern oyster (primary study species)
  • Crassostrea gigas: Pacific oyster (used for reference genome/annotation)
  • Perkinsus marinus: Dermo disease-causing parasite
  • Mytilus chilensis: Chilean blue mussel (methylation studies)

Related Pages: